EC220 - Game Theory
Slide Pack 10: Wrap-up
Costas Cavounidis
EC220 - Game Theory - Warwick - 2019/20
1 / 28
Summary: Game Types and Solution Concepts
2 / 28
Static Games of Complete Information
A static game of complete information has players, actions for
each player, the utility function of each player.
• Normal Form representation for finite games
• Simultaneous Play
• Pure and Mixed strategies
3 / 28
Solving Static Games of Complete Information
A Strictly Dominant Strategy for i is a strategy si such that
for all alternative strategies si0 and any profile of others’
strategies s−i we have that E[ui (si , s−i )] > E[ui (si0 , s−i )].
A strictly dominant strategy does better than the rest regardless
of what one’s opponents are doing.
Solution Concept: Strictly Dominant Strategy Equilibrium.
A strategy profile such that each player’s strategy is strictly
dominant, doing better than every other strategy regardless of
opponent actions.
4 / 28
Solving Static Games of Complete Information
A Strictly Dominated Strategy for i is a strategy si such that
there is an alternative strategy si0 such that for any profile of
others’ strategies s−i we have that E[ui (si , s−i )] < E[ui (si0 , s−i )].
Iterated Elimination of Strictly Dominated Strategies is the
process of removing strategies that are strictly dominated, then
repeating until no more strategies can be removed.
Solution Concept: Iterated Elimination Equilibrium. When
only one strategy profile survives IESDS, it is an Iterated
Elimination Equilibrium.
Problem: Often, neither a SDSE or an IEE exist.
5 / 28
Solving Static Games of Complete Information
A Best-Response of i to a profile of opponent strategies s−i is
a strategy si such that for any alternative strategy si0 , we have
that E[ui (si , s−i )] ≥ E[ui (si0 , s−i )].
Often, multiple best responses to a single s−i exist.
With differentiable utility functions and continuum pure action
sets, we can often find these by taking a First Order Condition,
and then checking the Second Order Condition.
Solution Concept: Nash Equilibrium. A profile of strategies s
such that for each i, we have that si is a best-response to s−i .
Every SDSE is a IEE. Every IEE is a NE.
6 / 28
Solving Static Games of Complete Information
A finite game has finitely many players and pure actions for
each player.
The Nash Existence Theorem states that every finite game
has at least one (potentially mixed) Nash Equilibrium.
In finite games, the underlining method will find every pure NE.
7 / 28
Dynamic Games of Complete Information
Static game of complete information plus information sets
• Extensive Form representation for finite games
• Strategies vs. actions
• Nodes characterize all previous actions
• Information sets are collections of nodes the acting player
can’t distinguish between.
• A strategy now gives an action at each own info set; it is a
complete contingent plan.
8 / 28
Solving Dynamic Games of Complete Information
Can put game in Normal Form and solve for Nash Equilibria.
BUT off-path play may be sequentially irrational.
• Equilibria with incredible threats etc.
Sequential Rationality: At each information set, a player plays
a best-response to the others’ strategies from then on.
9 / 28
Solving Dynamic Games of Complete Information
A Subgame is a node and all its descendant nodes, such that
no descendant’s information set has nodes outside these
descendant nodes.
A subgame is a self-contained game starting at some node of
the original game.
Solution Concept: Subgame Perfect Nash Equilibrium.
A Nash Equilibrium strategy profile that also induces a Nash
Equilibrium in every subgame.
10 / 28
Solving Dynamic Games of Complete Information
To find SPNEs, use Backwards Induction: start at the terminal
subgames of the game, identify NE there, and work your way
up the extensive form.
If a finite dynamic game of perfect information (all information
sets are singletons) never offers the same payoff to the same
player, it has a unique SPNE.
11 / 28
Static Games of Incomplete Information
Static game of complete information plus types
• Types can enter players’ utility functions
• Types represent differences in information about the game
• Strategies are now type-contingent
12 / 28
Solving Static Games of Incomplete Information
Harsanyi’s solution: add another parameter, a common prior
distribution over the players’ different types. "Nature"
Solution Concept: Bayes-Nash Equilibrium. A strategy
profile such that each type of each player best-responds to the
profile of opponent strategies after updating the common prior
based on their type.
In other words, a BNE is a strategy profile s such that for each
i, type θi of i, and alternative action ai0 , we have that
E[ui (si (θi ), s−i (θ−i ), θ)|θi ] ≥ E[ui (ai0 , s−i (θ−i ), θ)|θi ]
13 / 28
Dynamic Games of Incomplete Information
Dynamic game of complete information plus types
• Info sets now also used to track type information
14 / 28
Solving Dynamic Games of Incomplete Information
Can solve using BNE.
Problem: BNE may give sequentially irrational predictions
there, just as NE did.
But no proper subgames! So subgame perfection has no bite.
15 / 28
Solving Dynamic Games of Incomplete Information
What node does the active player thinks he’s at?
System of beliefs: For each information set, a probability
distribution over it’s nodes.
Solution Concept: Perfect Bayesian Equilibrium
A BNE in which each player is playing a best-response at each
of her information sets, given their beliefs over the nodes in that
information set. Beliefs restricted:
• Given by Bayes’ Rule using the prior and strats on-path
• Arbitrary off-path
By definition, every PBE is a BNE.
PBE gives Sequential Rationality in games of Incomplete Info!
16 / 28
The Cho-Kreps Intuitive Criterion
Many PBEs seem fishy!
We can eliminate some by applying the Intuitive Criterion.
Is there some song some type could sing, that it would sing if
doing so revealed it’s type? Is it the case that no other type
would want to sing that song no matter what this did to beliefs?
17 / 28
Nash-based Solution Concept Summary
SR can fail SR does not fail
SPNE: Every-
NE: Everyone is one is best-
Complete info
best-responding responding in
each subgame
BNE: Every PBE: Every
type is best- type is best-
Incomplete info responding given responding at
prior updated on each info set
type given beliefs
18 / 28
Summary: Applications
19 / 28
Finitely Repeated Games
A Finitely Repeated Game is a dynamic game G(δ, T ) in
which the same static "stage" game G is played for T periods,
with utility exponentially discounted by δ ≤ 1.
If there is a unique NE of the stage game, repeating it at each
history is the only SPNE of the finitely repeated game.
If not, an SPNE can have behavior different than stage-game
Nash play.
20 / 28
Infinitely Repeated Games
An Infinitely Repeated Game G(δ) is the same, only it
continues forever. Payoffs accrue along the way, and are
normalized by multiplication by (1 − δ).
Equivalence between per-period probability of breakdown
(1 − δ) and discounting at rate δ.
No final subgames, so can’t use Backwards Induction.
21 / 28
Infinitely Repeated Games
A vector of payoffs v is feasible if it lies in the convex hull of the
pure action payoff points.
If δ is high enough, some strategy profile (equilibrium or not)
results in payoffs v .
22 / 28
Infinitely Repeated Games
Playing a given Nash Equilibrium of the stage-game at every
history is an SPNE of the repeated game.
Nash Reversion strategies give a certain on-path play, and if a
deviation occurs, revert to playing a Nash Equilibrium of the
stage-game forevermore. Which NE they revert to can depend
on the history.
The Friedman Folk Theorem establishes that if a feasible
payoff vector v gives each player a payoff better than that
player’s worst stage-game Nash payoff, for high enough δ, v is
the payoff of some SPNE of G(δ).
Furthermore, such an SPNE can be constructed using Nash
Reversion.
23 / 28
Infinitely Repeated Games
Player i’s minmax payoff wi is the worst payoff she can get in
the stage-game G given that i is best responding.
wi = min max ui (ai , α−i )
α−i ∈×j6=i ∆Aj ai ∈Ai
The Fudenberg-Maskin Folk Theorem establishes that, if v is
a feasible payoff vector and vi > wi for all i, then v is the payoff
of some SPNE of G(δ).
24 / 28
Signaling Games
A signaling game is a dynamic game of incomplete information
with two players:
• A sender with multiple possible types who acts first; and
• A receiver with only one type that observes the sender’s
action (but not her type) and then acts.
We often refer to the sender’s action as the message.
A pooling equilibrium is a PBE in which all types of sender
send the same (potentially mixed) message.
A separating equilibrium is a PBE in which no two sender
types ever send the same message.
25 / 28
Information Aggregation
Information Aggregation refers to the extent to which, in
equilibrium, individual players’ information is combined into
outcomes.
In the sequential restaurant-choice game, in the PBE an
information cascade would eventually form and squelch
aggregation of all future information.
26 / 28
Reputation
Reputation models are dynamic settings in which players may
have incentives to convince other players they are of a certain
type - whether or not this is true!
In finitely-repeated settings, even if ‘behavioral’ types are
unlikely, very different results than with single type.
In infinitely repeated settings, patient players build reputation
even if ‘behavioral’ types are unlikely. Such a player does well
in every equilibrium, undoing folk-theorem-style results!
27 / 28
Thanks for Playing!
28 / 28